Arena: The Free Claude Code Skill That Makes 100 Claudes Fight for Your Best Answer (Setup Guide)

You ask Claude for something, you do not like the answer, and you ask again. And again. Arena is a free Claude Code skill built for exactly that moment. Instead of asking one Claude to try once more, it gives the same task to 100 sub-agents, each reasoning a different way, and makes them fight until one answer is left.

It is free, MIT licensed, needs no signup and no API key. The tokens it uses are yours, so read the cost section before running the full 100.

How Arena Works

1. Spawn. Every sub-agent gets the exact same task text, plus its own strategy card: a reasoning mode, a workflow and a strategy. The deck has 15 reasoning modes, 12 workflows and 12 strategies, which is 2,160 different cards.
2. Attack. Answers are paired off. Each side attacks the other's answer: what is wrong, what was missed, what input breaks it.
3. Defend. Each side answers the attacks it took, concedes or rebuts, and rewrites its answer.
4. Judge. A separate judge scores both revised answers on a written rubric. The loser is out.
5. Repeat. 100 becomes 50, then 25, 13, 7, 4, 2 and finally 1. You get that last answer, the attacks it survived and the number of rounds.

If you started from an answer you rejected, a last judge compares the winner against it blind and the skill tells you the score, even when your old answer wins.

Setup in 2 Steps

1. Open Claude Code and paste the repo link with "install this skill":
   https://github.com/Jakeschincariol/arena-skill Install this skill, then confirm /arena works.
2. Any time an answer feels weak, type /arena. By default it takes your last request as the task and the answer you did not like as the one to beat.

You can also copy the folder yourself with git clone https://github.com/Jakeschincariol/arena-skill.git and cp -r arena-skill/skills/arena ~/.claude/skills/. It needs Claude Code, because it spawns sub-agents, and Python 3.8 or newer. Nothing to pip install.

Flags You Will Actually Use

• --quick runs 16 competitors. The README calls it the everyday setting.
• --agents N sets the number of competitors. The default is 100.
• --seed S gives the same cards and the same bracket again. It does not give the same answers.

It Never Touches Your Files

Every sub-agent writes inside a .arena/ folder. The winning answer comes back to you, and using it is your call. Code changes come back as a diff or as full files inside the answer, and the skill asks before applying anything.

What It Costs

The skill is free. The tokens are not. Per the README, a full 100-agent run is about 595 sub-agent calls in 70 waves of 10, and the bill grows with the size of your task.

Use --quick for everyday work and keep the full 100 for the answer that matters.

The Fine Print (Read This Before You Trust It)

• "100 versions of Claude" means 100 sub-agents of the model you are running. They are not 100 different models. The strategy cards are what make them differ.
• "Best answer" means the one that survived every match. The judges are Claude too, scoring against a rubric you can read and change. It is the strongest answer the tournament found, not proof that it is right.
• Expect permission prompts unless you allow edits, because each sub-agent writes a file.
• Sub-agents cannot see your chat. Only the task file the skill writes. If a requirement is missing there, all 100 miss it.
• A vague task gives 100 flavours of vague. The README says so itself.

Arena was made by Jake Schincariol (opusjake.ai). Full credit to him for the skill.